<ul><li>The research article focuses on Bayesian ensembles and learning optimal combinations of Bayesian models in an online, continual learning setting.</li><li>The study reinterprets approaches like Bayesian model averaging (BMA) and Bayesian stacking using empirical Bayes lens, highlighting the limitations of BMA.</li><li>A new method called Online Bayesian Stacking (OBS) is introduced, which optimizes log-score over predictive distributions to adaptively combine Bayesian models, showing better performance in certain scenarios than online BMA.</li><li>The work establishes a connection between OBS and portfolio selection, offering efficient algorithms and regret analysis, providing guidance on when to prefer OBS over online BMA.</li></ul>

Bayesian Ensembling: Insights from Online Optimization and Empirical Bayes

Discover more